Infrared Weak And Small Target Detection Method Based On Deep Network And Its Application In UAV Detection

Infrared dim and small target detection technology is one of the key technologies of infrared imaging guidance,antimissile system,drone intrusion detection,and leak detection based on thermal imaging.Due to long-distance imaging and the large degradation of the target signal,the infrared image is susceptible to environmental radiation and sensor noise.The infrared image has a very low signal-to-noise ratio.The small target occupies very few pixels in the entire image and is difficult to rely on Size,texture or shape,and structural features are used for detection and judgment.We cannot accurately detect the targets based on the gray information of the target itself.Therefore,the infrared dim and small target detection method in complex scenes has always been a research difficulty in the field of target detection.It is of important practical significance to improve the correct detection rate of dim and small targets and reduce the false alarm rate.In view of the above difficulties,this paper proposes two infrared dim and small target detection models based on deep networks and applies them to the specific application of drone detection.The specific contents include:(1)An unsupervised end-to-end infrared small target detection model(CDAE)based on a denoising auto-encoder network and a convolutional neural network is proposed.We convert the problem of infrared dim and small target detection into a special "denoising" problem,that is,small targets are regarded as "noise points" in the infrared image;the CDAE model is divided into two processes of encoding and decoding,and is designed with a convolutional neural network Model architecture;In designing the loss function of the CDAE model,a structure loss and a perceptual loss are introduced to guide the CDAE model to learn the background features of the infrared image.(2)It is observed that the detection effect of the CDAE model is very dependent on the model's degree of learning the background features of the infrared image,which limits the detection ability of the CDAE model.In order to improve the detection method,this paper proposes an infrared dim and small target detection model(ISOS-GAN)based on generative adversarial networks.The ISOS-GAN model has two different generators that learn the background and the small target of the infrared image,and a discriminator allows the two to compete against each other to achieve Nash equilibrium.In the test phase,only a small target generator is needed.Compared with typical infrared dim and small target detection methods and CDAE model,the ISOS-GAN model can obtain better detection performance in the test sequence.(3)The above two infrared dim and small target detection methods based on deep network are further verified on drone detection tasks of infrared image.Due to the typical dim and small characteristics of drones,ordinary drones generally fly over cities,and interference in urban backgrounds is more complex and diverse.Therefore,our two models which directly applied on drone detection tasks would generate more false alarms in different extent.In order to overcome the background interference and reduce false alarms by learning as much as possible,we lead into the Sin GAN model to assist in generating simulation training samples.Sin GAN model only needs to read one frame infrared image to be detected,and can generate a large number of similar images with the same distribution.Experiments prove that retraining the ISOS-GAN model using the generated image as an auxiliary training set,not only achieves better performance on the UAV detection task.It also helps to detect dim and small infrared targets in other scenes.